NeurIPS 2024 Papers — Page 35
Conference on Neural Information Processing Systems · 4035 papers
Stability and Generalizability in SDE Diffusion Models with Measure-Preserving Dynamics
Weitong Zhang (Imperial College London), Bernhard Kainz (Imperial College London)
RestorationSuper ResolutionDiffusion modelImageBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: A measure-preserving dynamics-based SDE diffusion model D3GM is proposed to address stability and generalization issues in inverse problems.
Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation
Kaibo Zhang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
OptimizationAdversarial Attack
🎯 What it does: Analyzed and provided the stability and generalization performance of two-layer smooth activation function networks under adversarial training, proposing theoretical bounds based on Uniform Argument Stability.
Stability and Generalization of Asynchronous SGD: Sharper Bounds Beyond Lipschitz and Smoothness
Xiaoge Deng (National University of Defense Technology), Xicheng Lu (National University of Defense Technology)
OptimizationImageText
🎯 What it does: This paper studies the stability and generalization error of Asynchronous Stochastic Gradient Descent (ASGD), providing tighter bounds under weak assumptions.
Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective
Yongxin Zhu (University of Science and Technology of China), Lidong Bing (University of Science and Technology of China)
GenerationTransformerAuto EncoderImage
🎯 What it does: A discrete image tokenizer based on a stable latent space has been constructed, and a GPT-style causal Transformer generative model DiGIT has been trained in this space.
Stabilized Proximal-Point Methods for Federated Optimization
Xiaowen Jiang (Saarland University), Sebastian U Stich
OptimizationFederated LearningImageTabular
🎯 What it does: A new distributed proximal point method S-DANE and its accelerated version ACC-S-DANE are proposed for federated optimization.
Stabilizing Linear Passive-Aggressive Online Learning with Weighted Reservoir Sampling
Skyler Wu (Booz Allen Hamilton), James Holt (Laboratory for Physical Sciences)
Tabular
🎯 What it does: A WRS-Augmented Training (WAT) method is proposed that stabilizes any passive-aggressive online learning algorithm with single-pass training and without the need for a hold-out set.
Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks
Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)
RecognitionData-Centric LearningTransformerVision Language ModelMultimodality
🎯 What it does: In the continuous learning of pre-trained vision-language models, an anti-forgetting regularization method based on zero-shot prediction stability is introduced, along with the construction of an EMA-LoRA lightweight architecture.
Stable Minima Cannot Overfit in Univariate ReLU Networks: Generalization by Large Step Sizes
Dan Qiao (University of California San Diego), Yu-Xiang Wang (University of California San Diego)
OptimizationTabular
🎯 What it does: This study investigates the generalization performance of two-layer ReLU networks in one-dimensional nonparametric regression under noiseless labels, and proposes that during gradient descent with large step sizes, the network converges to a non-interpolating, smooth stable minimum.
Stable-Pose: Leveraging Transformers for Pose-Guided Text-to-Image Generation
Jiajun Wang (Technical University of Munich), Christian Wachinger (Technical University of Munich)
GenerationPose EstimationTransformerDiffusion modelImage
🎯 What it does: Developed the Stable-Pose adapter within the Stable Diffusion framework, achieving fine control of human poses through a coarse-to-fine hierarchical pose masking attention mechanism, enabling pose-guided text-to-image generation.
Stacking Your Transformers: A Closer Look at Model Growth for Efficient LLM Pre-Training
Wenyu Du (University of Hong Kong), Jie Fu (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This study systematically evaluates four types of atomic model growth operators, proposes the deep stacking operator G stack, and conducts comprehensive experiments on a unified LLM pre-training benchmark to verify its significant acceleration effect, further explores its scalability, and provides actionable growth timing and ratio criteria.
Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning
Hang Zhou (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
OptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: The Star-Agents framework is proposed, which generates diverse and high-quality instruction data through multi-agent collaboration and conducts dual model evaluation to enhance the instruction tuning effect of LLMs.
START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation
Jintao Guo (Nanjing University), Yang Gao (Nanjing University)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a domain generalization framework called START based on a state space model, which utilizes saliency-driven token-aware transformations to suppress domain features in the input dependency matrix, thereby enhancing the model's generalization ability to unseen domains.
State Chrono Representation for Enhancing Generalization in Reinforcement Learning
Jianda Chen (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)
Reinforcement LearningImage
🎯 What it does: A new state-time representation (SCR) method is proposed, which enhances the generalization ability of state representation in reinforcement learning by incorporating extensive temporal information into the update steps of dual-similarity metric learning.
State Space Models on Temporal Graphs: A First-Principles Study
Jintang Li (Sun Yat-sen University), Zibin Zheng (Shanghai Jiao Tong University)
CompressionComputational EfficiencyGraph Neural NetworkGraphTime Series
🎯 What it does: A novel state space model for temporal graphs, GRAPHSSM, is proposed to compress and predict the dynamic evolution of graphs.
State-free Reinforcement Learning
Mingyu Chen (Boston University), Xuezhou Zhang (Boston University)
Reinforcement LearningTabular
🎯 What it does: This paper studies reinforcement learning algorithms that do not require prior knowledge of the state space, proposing a state-free RL framework SF-RL that converts any RL algorithm into a state-free version.
Statistical and Geometrical properties of the Kernel Kullback-Leibler divergence
Anna Korba (CREST, ENSAE, IP Paris), Clémentine Chazal (CREST, ENSAE, IP Paris)
Optimization
🎯 What it does: This paper proposes a regularized variant of the original Kernel Kullback-Leibler (KKL) divergence, providing its closed-form expression and gradient, and implements Wasserstein gradient flow optimization on discrete measures; theoretical analysis gives bias bounds, finite sample bounds, and convergence relations with the original KKL; the effectiveness of the method is subsequently validated on synthetic datasets.
Statistical Efficiency of Distributional Temporal Difference Learning
Yang Peng (Peking University), Zhihua Zhang (Peking University)
Reinforcement LearningTabular
🎯 What it does: This paper conducts a theoretical analysis of the statistical efficiency of distributional temporal difference learning (distributional TD) under finite sample conditions, proposes non-parametric distributional TD (NTD), and provides an upper bound on the number of iterations under the 1-Wasserstein metric. It subsequently proves that the more commonly used category TD (CTD) can also achieve the same non-asymptotic convergence rate.
Statistical Estimation in the Spiked Tensor Model via the Quantum Approximate Optimization Algorithm
Leo Zhou (California Institute of Technology), Song Mei (University of California)
OptimizationPhysics Related
🎯 What it does: This paper provides a theoretical and numerical analysis of the weak recovery performance of the Quantum Approximate Optimization Algorithm (QAOA) on the spiked tensor model.
Statistical Multicriteria Benchmarking via the GSD-Front
Christoph Jansen (Lancaster University), Thomas Augustin (Ludwig-Maximilians-Universität München)
ClassificationOptimizationTabularBenchmark
🎯 What it does: This paper proposes a multi-criteria classifier comparison framework based on the Generalized Stochastic Dominance (GSD) frontier, which can simultaneously consider multiple quality metrics and provide statistical significance judgments.
Statistical-Computational Trade-offs for Density Estimation
Anders Aamand (University of Copenhagen), Haike Xu (Massachusetts Institute of Technology)
OptimizationComputational EfficiencyTabular
🎯 What it does: The study addresses the density estimation problem using sublinear sample sizes in discrete domains, providing a lower bound for the statistical-computational trade-off and offering a matching upper bound algorithm for a special case of semi-uniform distribution.
Stealth edits to large language models
Oliver Sutton, Ivan Y Tyukin
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes and implements the method of 'Stealth Edits', which can precisely correct the hallucination outputs of language models under specific prompts through fine-grained weight updates without retraining the model; it also reveals the model's susceptibility to malicious stealth attacks.
StepbaQ: Stepping backward as Correction for Quantized Diffusion Models
Yi-Chung Chen (MediaTek), Jing-Ren Chen (MediaTek)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A compensation method called StepbaQ is proposed to correct the accumulation of quantization errors in quantized diffusion models, which can improve generation quality without modifying the quantization settings.
Stepping Forward on the Last Mile
Chen Feng (Qualcomm), Andrew Zou Li (University of Toronto)
OptimizationImageAudio
🎯 What it does: Proposes the use of fixed-point forward gradients for model adaptive training on low-power edge devices, avoiding the high memory consumption of traditional backpropagation;
Stepping on the Edge: Curvature Aware Learning Rate Tuners
Vincent Roulet (Google DeepMind), Fabian Pedregosa (Google DeepMind)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper analyzes the closed-loop feedback between learning rate schedulers and curvature (maximum Hessian eigenvalue), revealing that traditional greedy or line search schedulers underestimate the stability boundary during full-batch training, leading to a decrease in learning rate and an increase in curvature, which ultimately harms long-term performance. It then proposes Curvature Dynamics Aware Tuning (CDAT), which adaptively maintains the learning rate close to or slightly above the stability boundary of curvature, thereby achieving better training convergence and built-in warm-up behavior in both full-batch and mini-batch scenarios.
Stepwise Alignment for Constrained Language Model Policy Optimization
Akifumi Wachi (LY Corporation), Youhei Akimoto (University of Tsukuba)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a stepwise alignment approach for optimizing constrained language models (SACPO), which first aligns reward metrics and then aligns safety metrics, thereby achieving dual alignment of human values and safety constraints.
STL: Still Tricky Logic (for System Validation, Even When Showing Your Work)
Isabelle Hurley (Lincoln Laboratory Massachusetts Institute of Technology), Ho Chit Siu (Lincoln Laboratory Massachusetts Institute of Technology)
Robotic IntelligenceReinforcement LearningTime Series
🎯 What it does: The study helps humans verify whether robot motion strategies based on Signal Temporal Logic (STL) meet expected goals through active learning and real-time feedback methods.
Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution
Ian Connick Covert, Tatsunori Hashimoto (Stanford University)
Explainability and InterpretabilityComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkImageTabular
🎯 What it does: A random amortization method is proposed to accelerate feature and data attribution tasks by training models with noisy labels.
Stochastic Concept Bottleneck Models
Moritz Vandenhirtz (ETH Zurich), Julia E Vogt
ClassificationOptimizationImageTabular
🎯 What it does: A new concept bottleneck model (SCBM) is proposed, which explicitly models the correlation between concepts using a multivariate normal distribution, and based on this, designs an intervention strategy based on confidence intervals.
Stochastic contextual bandits with graph feedback: from independence number to MAS number
Yuxiao Wen (New York University), Zhengyuan Zhou (New York University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper studies the random context Bandit problem with feedback graphs, proposing new lower and upper bounds along with corresponding algorithms.
Stochastic Extragradient with Flip-Flop Shuffling & Anchoring: Provable Improvements
Jiseok Chae (KAIST), Donghwan Kim (KAIST)
Optimization
🎯 What it does: A new stochastic external gradient method SEG-FFA is proposed, which combines flipping-reverse shuffling and anchoring to achieve convergence and improve the convergence rate for unconstrained convex-concave and strongly convex-strongly concave problems.
Stochastic Kernel Regularisation Improves Generalisation in Deep Kernel Machines
Edward Milsom (University of Bristol), Laurence Aitchison (University of Bristol)
ClassificationOptimizationConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: This paper introduces random kernel regularization on deep kernel machines, enhancing their generalization ability for image classification tasks.
Stochastic Newton Proximal Extragradient Method
Ruichen Jiang (University of Texas at Austin), Aryan Mokhtari (University of Texas at Austin)
Optimization
🎯 What it does: This study investigates the use of the Stochastic Newton Proximal Extragradient method (SNPE) to solve strongly convex smooth optimization problems under the condition of only having access to gradients and noisy Hessians, and provides convergence analysis for both uniform and weighted Hessian averaging schemes.
Stochastic Optimal Control and Estimation with Multiplicative and Internal Noise
Francesco Damiani (Pompeu Fabra University), Rubén Moreno-Bote (Pompeu Fabra University)
OptimizationTime SeriesStochastic Differential Equation
🎯 What it does: For the stochastic optimal control problem with multiplicative noise and internal noise, an unbiased estimation hypothesis error correction gradient descent and analytical FPOMP algorithm are proposed to solve the optimal control and filter using linear control laws.
Stochastic Optimal Control for Diffusion Bridges in Function Spaces
Byoungwoo Park (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisOptimizationTime SeriesStochastic Differential Equation
🎯 What it does: Construct a stochastic optimal control theory for diffusion bridges in infinite-dimensional Hilbert spaces, and based on this, design learning algorithms for bridge matching and Bayesian inference, solving the problem of density without explicit form caused by the lack of equivalent Lebesgue measure;
Stochastic Optimal Control Matching
Carles Domingo-Enrich (New York University), Ricky T. Q. Chen (Meta)
OptimizationReinforcement LearningStochastic Differential Equation
🎯 What it does: Proposed the SOCM method, which uses least squares matching vector fields to learn stochastic optimal control.
Stochastic Optimization Algorithms for Instrumental Variable Regression with Streaming Data
Xuxing Chen (University of California), Krishna Balasubramanian
OptimizationTabularTime Series
🎯 What it does: Two pure online algorithms (TOSG-IVaR and OTSG-IVaR) are proposed to directly solve the conditional random optimization problem of IV regression, without the need for matrix inversion or mini-batch, enabling least squares IV regression in a streaming data environment.
Stochastic Optimization Schemes for Performative Prediction with Nonconvex Loss
Qiang LI, Hoi To Wai
OptimizationTabular
🎯 What it does: This paper studies the optimization of non-convex loss using Stochastic Gradient Descent (SGD) under decision-related data distributions, analyzing the convergence of greedy deployment SGD-GD and lazy deployment schemes;
Stochastic Taylor Derivative Estimator: Efficient amortization for arbitrary differential operators
Zekun Shi (National University of Singapore), Kenji Kawaguchi (National University of Singapore)
OptimizationComputational EfficiencyTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A stochastic Taylor derivative estimator (STDE) based on infinitesimal Taylor mode automatic differentiation is proposed, which can efficiently randomize differential operators of any order and any dimension;
Stochastic Zeroth-Order Optimization under Strongly Convexity and Lipschitz Hessian: Minimax Sample Complexity
Qian Yu (University of California), Jason D. Lee (Princeton University)
Optimization
🎯 What it does: This study investigates stochastic zeroth-order optimization under second-order smooth and strongly convex objective functions, providing upper and lower bounds for optimal simple regret, and proposing an algorithm that can achieve this performance.
STONE: A Submodular Optimization Framework for Active 3D Object Detection
RUIYU MAO, Yunhui Guo (University of Texas at Dallas)
Object DetectionAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: A submodular optimization-based active 3D object detection framework called STONE is proposed, which significantly reduces labeling costs using a two-stage subset selection strategy.
Stopping Bayesian Optimization with Probabilistic Regret Bounds
James T. Wilson (Morgan Stanley)
OptimizationHyperparameter SearchTabularSequential
🎯 What it does: This paper proposes a Probability Regret Bound (PRB) stopping rule based on Bayesian optimization, which uses a Gaussian process model to estimate the probability that the current point satisfies the ε-optimal condition and stops the search once a threshold is reached.
StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation
Yupeng Zhou (Nankai University), Qibin Hou (Nankai University)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: We propose StoryDiffusion, which generates image sequences with consistent identity and attire while maintaining text controllability, and transforms these images into coherent videos through a semantic motion predictor.
Strategic Linear Contextual Bandits
Thomas Kleine Buening (Alan Turing Institute), Haifeng Xu (University of Chicago)
Reinforcement Learning
🎯 What it does: Designed and analyzed mechanisms for the strategic linear contextual bandit problem (Greedy Grim Trigger Mechanism and Optimistic Grim Trigger Mechanism), which incentivize self-interested arms to honestly report contexts while maximizing their own pull counts, thereby achieving approximately optimal low regret in a strategic environment.
Strategic Littlestone Dimension: Improved Bounds on Online Strategic Classification
Saba Ahmadi (Toyota Technological Institute at Chicago), Hanrui Zhang (Chinese University of Hong Kong)
ClassificationOptimization
🎯 What it does: This study addresses the issue of intelligent agents in online binary classification striving for positive class labels by modifying features, proposing a new combinatorial complexity measure called 'Strategic Littlestone Dimension (SLdim)'. It provides optimal error/regret bounds in both realizable and unrealizable scenarios, and also examines the case of unknown manipulation graphs, offering corresponding error/regret analyses.
Strategic Multi-Armed Bandit Problems Under Debt-Free Reporting
Ahmed Ben Yahmed (Criteo AI Lab), Vianney Perchet (Criteo AI Lab)
OptimizationReinforcement LearningFinance Related
🎯 What it does: A strategic multi-armed bandit model under the condition of debt-free reporting is proposed, and an incentive-based sequential elimination algorithm S-SE is designed, allowing each arm to report honestly with a dominant strategy, thereby achieving near-optimal rewards and obtaining suboptimal average returns of logarithmic or square root order under this equilibrium.
StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving
Chang Gao (Chinese University of Hong Kong), Wai Lam (Chinese University of Hong Kong)
OptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper proposes the StrategyLLM framework, which utilizes large language models to automatically generate, execute, optimize, and evaluate task-level strategies, thereby constructing general and consistent few-shot prompts to enhance the performance of various reasoning tasks.
Stratified Prediction-Powered Inference for Effective Hybrid Evaluation of Language Models
Adam Fisch (Google DeepMind), William W. Cohen (Google DeepMind)
Large Language ModelText
🎯 What it does: A hybrid method combining human annotation and automatic evaluators in large language model evaluation is proposed—Stratified Prediction-Powered Inference (StratPPI), which improves the PPI method by performing stratified sampling and weighted estimation on pre-defined subdomains.
StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences
Shangkun Sun (Peking University), Wei Gao (Peking University)
Computational EfficiencyTransformerOptical FlowVideo
🎯 What it does: A multi-frame optical flow estimation framework for video streams, StreamFlow, is proposed, which can predict continuous unidirectional optical flow in a single forward pass while significantly reducing redundant computations.
Streaming Bayes GFlowNets
Tiago Silva, Diego Mesquita (Getulio Vargas Foundation)
Flow-based ModelTabular
🎯 What it does: The SB-GFlowNet method for streaming Bayesian inference in discrete parameter spaces is proposed, which can update the posterior distribution with new data without retraining;
Streaming Long Video Understanding with Large Language Models
Rui Qian (Chinese University of Hong Kong), Jiaqi Wang (Shanghai AI Laboratory)
RecognitionRetrievalOptimizationTransformerLarge Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: Proposes VideoStreaming, which utilizes memory propagation for streaming encoding and adaptive memory selection to achieve long video understanding.
StreamingDialogue: Prolonged Dialogue Learning via Long Context Compression with Minimal Losses
Jia-Nan Li (Renmin University of China), Rui Yan (Renmin University of China)
CompressionComputational EfficiencyTransformerText
🎯 What it does: By treating the End-of-Utterance (EoU) delimiters in conversations as 'conversation attention absorption points', long dialogue histories are compressed, only caching these attention absorption points, thereby reducing computational load and memory usage, supporting continuous dialogues of over 200K statements.
Stress-Testing Capability Elicitation With Password-Locked Models
Ryan Greenblatt (Redwood Research), David Krueger (University of Cambridge)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper studies a type of LLM called password-locked models, which only exhibit hidden capabilities when a specific password appears in the input; otherwise, their performance is weakened. The model is used to stress-test capability activation methods based on fine-tuning.
Structural Inference of Dynamical Systems with Conjoined State Space Models
Aoran Wang (University of Luxembourg), Jun Pang (University of Luxembourg)
Flow-based ModelTime SeriesSequentialBenchmark
🎯 What it does: Proposed and implemented the SICSM (Structural Inference with Conjoined State Space Models) framework for inferring the hidden structure of systems from irregularly sampled and partially observed dynamic trajectories.
Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis
Rui Peng (Peking University), Ronggang Wang (Peking University)
GenerationData SynthesisOptimizationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: A new perspective synthesis method based on 3D Gaussian projection with few views (SCGaussian) is proposed, which matches prior learned 3D consistent scene structures.
Structured flexibility in recurrent neural networks via neuromodulation
Julia C Costacurta, Scott Linderman
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: A dynamic weight-adjustable neural modulation recurrent network (NM-RNN) was constructed and evaluated, which can be achieved through low-rank recurrent networks and neural modulation subnetworks, and applied to time measurement, multi-task learning, and long-term dependency tasks.
Structured Learning of Compositional Sequential Interventions
Jialin Yu (University College London), Ricardo Silva (University College London)
Recommendation SystemExplainability and InterpretabilityRecurrent Neural NetworkAuto EncoderSequential
🎯 What it does: This paper proposes an interpretable structured model for predicting behavioral changes under unknown combinations of sequential interventions (such as multi-time point policies or recommendations), and provides its identifiability theory and learnable algorithms.
Structured Matrix Basis for Multivariate Time Series Forecasting with Interpretable Dynamics
Xiaodan Chen (Harbin Institute of Technology), Zhijun Li (Harbin Institute of Technology)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTime Series
🎯 What it does: A multivariate time series forecasting model SUMBA based on structured matrix bases is proposed, which reduces variance and enhances interpretability by directly parameterizing dynamic spatial structures.
Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling
Jingwei Zhao (National University of Singapore), Ye Wang (National University of Singapore)
GenerationData SynthesisTransformerAuto EncoderAudio
🎯 What it does: A two-stage music accompaniment arrangement system is proposed, which first converts the main melody and chords into piano accompaniment, and then uses a style prior model to convert the piano accompaniment into a multi-track full orchestra arrangement.
Structured Unrestricted-Rank Matrices for Parameter Efficient Finetuning
Arijit Sehanobish (Independent), Snigdha Chaturvedi (University of North Carolina Chapel Hill)
SegmentationOptimizationTransformerSupervised Fine-TuningImageTextBiomedical Data
🎯 What it does: A parameter-efficient fine-tuning framework based on Structured Unranked Matrices (SURM) is proposed to replace traditional LoRA and Adapter, enhancing the fine-tuning effect of Transformer models while maintaining an extremely low parameter count.
Style Adaptation and Uncertainty Estimation for Multi-Source Blended-Target Domain Adaptation
Yuwu Lu (South China Normal University), Xue Hu (South China Normal University)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: A multi-source mixed domain adaptation (MBDA) method SAUE is proposed, which utilizes feature information from multiple source domains and achieves transfer learning for unlabeled mixed target domains through style adaptation and uncertainty estimation.
Stylus: Automatic Adapter Selection for Diffusion Models
Michael Luo (University of California Berkeley), Ion Stoica (University of California Berkeley)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A system named Stylus has been designed to retrieve and automatically combine the most relevant adapters from hundreds of thousands of LoRA adapters, enhancing the visual quality and text consistency of text-to-image generation.
Sub-optimal Experts mitigate Ambiguity in Inverse Reinforcement Learning
Riccardo Poiani (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
Reinforcement Learning
🎯 What it does: This study investigates the construction and estimation of feasible reward sets in Inverse Reinforcement Learning (IRL) when there are multiple suboptimal experts (with known relative optimality bounds).
SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning
Jiying Zhang (International Digital Economy Academy), Yu Li (International Digital Economy Academy)
Representation LearningDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: This paper proposes SubgDiff, a molecular representation learning method that incorporates subgraph information into diffusion models.
Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning
Yanting Miao (University of Waterloo), Yeqing Li (Google)
GenerationData SynthesisOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: Proposes the λ-Harmonic reward function and Reward Preference Optimization (RPO) scheme for lightweight fine-tuning of pre-trained diffusion models with a small number of reference images to achieve theme-driven text-to-image generation.
Subsurface Scattering for Gaussian Splatting
Jan-Niklas Dihlmann (University of Tübingen), Hendrik Lensch
GenerationData SynthesisComputational EfficiencyGaussian SplattingImage
🎯 What it does: This paper proposes a 3D Gaussian distribution rendering framework that combines surface lighting with an implicit scattering model, capable of real-time reconstruction and relighting of translucent objects from multi-view, multi-light images.
Subwords as Skills: Tokenization for Sparse-Reward Reinforcement Learning
David Yunis (Toyota Technological Institute at Chicago), Matthew Walter (Toyota Technological Institute at Chicago)
Reinforcement Learning
🎯 What it does: Quickly extract limited interpretable skills from demonstration actions using BPE subword tokenization, and use these skills as a discrete action space for sparse reward reinforcement learning.
Suitable is the Best: Task-Oriented Knowledge Fusion in Vulnerability Detection
Jingjing Wang (Institute of Systems Engineering, Academy of Military Sciences, PLA), Xiaohui Kuang (Institute of Systems Engineering, Academy of Military Sciences, PLA)
Graph Neural NetworkGraph
🎯 What it does: A knowledge fusion-based graph neural network vulnerability detection method, KF-GVD, is proposed, which can adaptively learn in specific task scenarios while maintaining generalization ability for the source task.
Super Consistency of Neural Network Landscapes and Learning Rate Transfer
Lorenzo Noci (ETH Zurich), Antonio Orvieto
TransformerSupervised Fine-TuningImageText
🎯 What it does: This study investigates the characteristics of the loss landscape of neural networks at different scales (width and depth) and proposes and validates the concept of 'Super Consistency';
SuperDeepFool: a new fast and accurate minimal adversarial attack
Alireza Abdolahpourrostam, Seyed-Mohsen Moosavi-Dezfooli (Apple)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: The SuperDeepFool (SDF) algorithm is proposed for efficiently solving the minimum ℓ2 adversarial perturbation in a white-box scenario, and it is applied to model adversarial training and evaluation.
Superposed Decoding: Multiple Generations from a Single Autoregressive Inference Pass
Ethan Shen (University of Washington), Aditya Kusupati (University of Washington)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the Superposed Decoding algorithm, which utilizes token embedding superposition and n-gram interpolation to generate k drafts in a single autoregressive inference, significantly reducing computational costs.
Supervised Kernel Thinning
Albert Gong (Cornell University), Raaz Dwivedi (Cornell University)
CompressionComputational EfficiencyTabular
🎯 What it does: A supervised compression method based on Kernel Thinning is proposed to accelerate the training and inference of Nadaraya-Watson and Kernel Ridge Regression.
SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition
Feng Lu (Tsinghua University), Chun Yuan (Tsinghua University)
RecognitionRetrievalDomain AdaptationTransformerImage
🎯 What it does: A novel global descriptor called SuperVLAD is proposed, which achieves more compact and robust visual location recognition features by removing cluster centers and using very few clusters. Additionally, a 1-Cluster VLAD is designed to obtain extremely low-dimensional descriptors.
Suppress Content Shift: Better Diffusion Features via Off-the-Shelf Generation Techniques
Benyuan Meng (Institute of Information Engineering), Qingming Huang (University of Chinese Academy of Sciences)
SegmentationGenerationDiffusion modelImage
🎯 What it does: This study investigates the phenomenon of content shift present in diffusion features and proposes to suppress this phenomenon using existing generative control techniques (such as ControlNet, LoRA, and fine-grained prompts) through the GATE framework, thereby enhancing feature quality.
Supra-Laplacian Encoding for Transformer on Dynamic Graphs
Yannis Karmim (Conservatoire National des Arts et Métiers), Nicolas THOME
Graph Neural NetworkTransformerGraphTime Series
🎯 What it does: Developed the SLATE model, which combines supra-Laplacian encoding with a fully connected Transformer to address dynamic link prediction.
SureMap: Simultaneous mean estimation for single-task and multi-task disaggregated evaluation
Mikhail Khodak (Princeton University), Miroslav Dudík (Microsoft Research)
TabularAudio
🎯 What it does: A hierarchical Gaussian prior method based on MAP estimation and SURE tuning, called SureMap, is proposed for the decomposition evaluation of subgroup performance in both single-task and multi-task scenarios.
Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling
Shuaipeng Li (Tencent Hunyuan), Di Wang (Tencent Hunyuan)
OptimizationHyperparameter SearchConvolutional Neural NetworkTransformerImageText
🎯 What it does: This study investigates the scaling relationship between the learning rate and batch size of Adam-style optimizers, theoretically deriving the 'fluctuation' pattern where the learning rate first increases and then decreases with batch size, and validating this pattern through large-scale experiments.
SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors
Vijay Lingam (University of Texas at Austin), sujay sanghavi
TransformerSupervised Fine-TuningImageTextBenchmark
🎯 What it does: For parameter-efficient fine-tuning (PEFT) of large-scale pre-trained models, a method called SVFT is proposed, which utilizes a sparse weighted combination of the singular vectors of the weight matrix to update the weights.
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
John Yang (Princeton University), Ofir Press (Princeton University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an Agent-Computer Interface (ACI) specifically designed for language model agents (LM agents) and builds the SWE-agent on this basis, which can automatically write, search, edit, and execute code in real software engineering tasks.
Swift Sampler: Efficient Learning of Sampler by 10 Parameters
Jiawei Yao (University of Washington), Canran Xiao (Central South University)
OptimizationHyperparameter SearchReinforcement LearningImage
🎯 What it does: This paper proposes an algorithm for an automated search data sampler (sampler) - Swift Sampler (SS), which accelerates model convergence and improves final performance by assigning sampling probabilities to each sample during training.
SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention
Róbert Csordás (Stanford University), Jürgen Schmidhuber (AI Initiative KAUST)
OptimizationComputational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: This paper proposes SwitchHead, a Transformer architecture that applies Mixture-of-Experts to the attention layer, significantly reducing computational and memory requirements while maintaining language modeling performance comparable to dense models.
SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents
Niels Mündler (ETH Zurich), Martin Vechev (ETH Zurich)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: The SWT-Bench benchmark is proposed to evaluate the automatic generation of unit tests by LLM code agents and the reproduction of GitHub issues, and to systematically assess different methods.
Symbolic Regression with a Learned Concept Library
Arya Grayeli (University of Texas at Austin), Swarat Chaudhuri (University of Texas at Austin)
Large Language ModelPrompt EngineeringTextPhysics Related
🎯 What it does: A concept library learning and evolution framework called LASR is proposed, which generates abstract textual concepts through zero-shot prompting and guides genetic algorithm searches, significantly improving performance in Feynman equations, synthesis tasks, and LLM scaling law discovery.
SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization
Qian Chen (Chinese University of Hong Kong), Tsung-Hui Chang (Chinese University of Hong Kong)
OptimizationGraph Neural NetworkGraph
🎯 What it does: The SymILO framework is proposed, which improves optimal solution prediction by introducing the symmetry of ILP in supervised learning.
Symmetric Linear Bandits with Hidden Symmetry
Nam Phuong Tran (University of Warwick), Long Tran-Thanh (University of Warwick)
OptimizationReinforcement Learning from Human FeedbackGaussian SplattingTabular
🎯 What it does: This paper studies the hidden symmetric subgroup G in high-dimensional linear stochastic bandits and proposes the Explore-Models-then-Commit (EMC) algorithm to achieve efficient learning under unknown symmetry.
Symmetries in Overparametrized Neural Networks: A Mean Field View
Javier Maass Martínez (Center for Mathematical Modeling University of Chile), Joaquin Fontbona
Stochastic Differential Equation
🎯 What it does: This paper develops a mean-field perspective on the learning dynamics of over-parameterized artificial neural networks, studying the distribution symmetry of data under the action of a general compact group G. A class of generalized shallow neural networks is considered, which are jointly trained using stochastic gradient descent (SGD) and possible symmetry-utilizing techniques such as data augmentation, feature averaging, or equivariant architectures.
Symmetry Discovery Beyond Affine Transformations
Ben Shaw (Utah State University), Kevin R. Moon (Utah State University)
Generative Adversarial NetworkTime Series
🎯 What it does: A method for detecting continuous symmetries based on vector fields is studied, which can discover symmetries beyond affine transformations and use them to construct invariant feature spaces.
Symmetry-Informed Governing Equation Discovery
Jianke Yang (University of California San Diego), Rose Yu (University of California San Diego)
OptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes embedding time-invariant symmetry (Lie point symmetry) into the symbolic equation discovery process. It first derives the invariance constraints of ODE flow mappings, and then explicitly solves linear symmetry constraints or incorporates symmetry regularization in methods such as sparse regression (SINDy) and genetic programming (GP), further enhancing the success rate of discovery and model simplicity under noisy data.
Synatra: Turning Indirect Knowledge into Direct Demonstrations for Digital Agents at Scale
Tianyue Ou (Carnegie Mellon University), Shuyan Zhou (Carnegie Mellon University)
Data SynthesisRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the Synatra method, which utilizes existing indirect knowledge (such as tutorials and random web pages) to generate direct demonstration data through large language models for training digital agents.
SyncTweedies: A General Generative Framework Based on Synchronized Diffusions
Jaihoon Kim (Korea Advanced Institute of Science and Technology), Minhyuk Sung (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: A general generative framework based on synchronous diffusion, SyncTweedies, is proposed, which can generate various visual content (such as ambiguous images, panoramic images, 3D mesh textures, and 3D Gaussian sphere textures) using pre-trained image diffusion models without additional fine-tuning.
SyncVIS: Synchronized Video Instance Segmentation
rongkun Zheng, Hengshuang Zhao (University of Hong Kong)
Object DetectionSegmentationTransformerVideo
🎯 What it does: A synchronous video instance segmentation framework SyncVIS is proposed, which enhances the modeling capability for complex videos by synchronously processing video-level and frame-level queries.
Synergistic Dual Spatial-aware Generation of Image-to-text and Text-to-image
Yu Zhao (Tianjin University), Jianguo Wei (Harbin Institute of Technology)
Image TranslationGenerationData SynthesisGraph Neural NetworkTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a bidirectional learning framework that jointly addresses the spatial perception tasks of image-to-text (SI2T) and text-to-image (ST2I);
Synthesize, Partition, then Adapt: Eliciting Diverse Samples from Foundation Models
Yeming Wen (University of Texas at Austin), Swarat Chaudhuri (University of Texas at Austin)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A framework called SPA is proposed, which utilizes synthetic data, data partitioning, and model adaptation to generate diverse responses while maintaining quality.
Synthetic Programming Elicitation for Text-to-Code in Very Low-Resource Programming and Formal Languages
Federico Mora (University of California Berkeley), Sanjit A. Seshia (University of California Berkeley)
GenerationAI Code AssistantLarge Language ModelPrompt EngineeringText
🎯 What it does: A text-to-code generation method called SPEAC is proposed for very low-resource programming languages, utilizing intermediate languages and automatic repair to achieve syntactic correctness.
T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback
Jiachen Li (University of California Santa Barbara), William Yang Wang (University of California Santa Barbara)
GenerationKnowledge DistillationReinforcement LearningVideoText
🎯 What it does: T2V-Turbo is proposed, which utilizes mixed reward feedback to enhance the generation quality of video consistency models during the consistency distillation process, achieving high-quality video generation in just 4-8 steps.
TabEBM: A Tabular Data Augmentation Method with Distinct Class-Specific Energy-Based Models
Andrei Margeloiu (University of Cambridge), Mateja Jamnik (University of Cambridge)
ClassificationData SynthesisGenerative Adversarial NetworkTabularStochastic Differential Equation
🎯 What it does: This paper proposes TabEBM, an energy model-based table data augmentation method that generates synthetic samples by constructing independent EBMs for each category to enhance the performance of downstream classifiers.
TableRAG: Million-Token Table Understanding with Language Models
Si-An Chen (National Taiwan University), Tomas Pfister (Google Cloud AI Research)
GenerationRetrievalTransformerLarge Language ModelTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the TableRAG framework, which utilizes retrieval-augmented generation (RAG) to achieve efficient reasoning of language models on large tables, and establishes a benchmark for table understanding at the million-level.
TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy
Weichao Zhao (University of Science and Technology of China), Can Huang (ByteDance)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelTabular
🎯 What it does: TabPedia is proposed, a unified visual table understanding large visual language model that can simultaneously perform table detection, structure recognition, location querying, and question-answering tasks.
Tackling Uncertain Correspondences for Multi-Modal Entity Alignment
Liyi Chen (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology)
TransformerLarge Language ModelAuto EncoderContrastive LearningMultimodality
🎯 What it does: A new model TMEA is proposed for entity alignment in multimodal knowledge graphs, focusing on addressing the uncertainty of correspondence both between different modalities and within the same modality.
Tactile DreamFusion: Exploiting Tactile Sensing for 3D Generation
Ruihan Gao (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationData SynthesisDiffusion modelMultimodality
🎯 What it does: This paper proposes the use of tactile perception to enhance the geometric details of 3D asset generation, combining visual and tactile textures in a 3D texture field for joint optimization.
TAIA: Large Language Models are Out-of-Distribution Data Learners
Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The TAIA method is proposed, which trains all parameters during LLM fine-tuning but retains only the updated attention parameters during inference, reducing reliance on high-quality domain data and enhancing robustness to OOD data.